{"id":"1bb40afc-63ab-4e68-90d1-6d58c93c9924","arxiv_id":"2412.20666","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"An unsupervised method detects vanishing points by fitting implicit lines through recurring feature correspondences and combining them with explicit lines via weighted RANSAC, outperforming classical and supervised methods on its new recurring-pattern benchmarks.","lead":"This paper introduces an unsupervised vanishing point detection method that uses repeated patterns in an image, not just straight lines, to find where parallel 3D lines meet. It also releases two new benchmark datasets and shows the approach competes with supervised deep learning methods on real images.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The RPVP benchmarks are built from the method's own prior—objects placed on L(t)=O+tD and GT drawn through recurring instances—so the headline parity claim may not transfer to arbitrary images with repeated patterns. Independent tests are needed.","rationale":"The reader's weakest_assumption identifies the same load-bearing concern: the evaluation is circular with respect to the method's core geometric prior. I agree that this is the most important threat to the central claim, and I add the post-hoc p-value detail (0.003 treated as insignificant against a threshold of 0.001) as a secondary weakness on TMM17-Test. The paper still has value: the method is unsupervised, the datasets are new, and on these curated RP images R-VPD appears to work. But the advertised 'on par with supervised learning on real-world images' is only demonstrated on images selected and annotated to satisfy the method's assumption. Therefore the original CONDITIONAL verdict remains appropriate; no change is needed, but the condition should be an independent evaluation on scenes not built from the receding-line prior.","tokens_in":15958,"tokens_out":5248,"duration_ms":59012,"concrete_test":"Run R-VPD and NeurVPS on an independent VPD benchmark with GT from explicit line intersections (e.g., YorkUrban or Eurasian Cities), and on a negative control built from RPVP-Synthetic scenes with identical objects arranged fronto-parallel or randomly jittered perpendicular to the receding line. If R-VPD's advantage over supervised methods disappears, or if it produces confident false VPs on the control, the headline should be restricted to 'scenes where RP instances lie on receding 3D lines.' A cheaper check: flip the scale-orientation convention (smaller-to-larger) and rerun Tables 1–3; if accuracy is unchanged, the scale prior is not the active mechanism being evaluated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim ('unsupervised approach outperforms all methods on synthetic, outperforms classical, and is on par with supervised learning on real-world images') depends on the RPVP datasets being representative of the real-world condition the method exploits. Those datasets are constructed from exactly the geometric prior used by R-VPD. In Appendix B.1.1, synthetic objects are placed along L(t)=O+tD and oriented toward the VP, so every recurring pattern instance forms a receding line with monotonically shrinking SIFT scale. For RPVP-Real, Section 4.1.2/B.2 states that GT VPs are computed as intersections of manually drawn lines passing through recurring patterns, and Section 4.1.3 notes over 70% overlap with TMM17, indicating that the selected images are ones where recurring instances already lie on receding lines. The method's core inference in Section 3.4 is 'vectors from larger to smaller scale features regarded as pointing toward the vanishing point'—i.e., exactly this assumption. Thus Tables 1–3 compare methods on data generated from R-VPD's prior. This is not an internal inconsistency, and R-VPD may genuinely work on such scenes; but the benchmark cannot establish the advertised generalization to arbitrary recurring-pattern images, such as fronto-parallel brick or tile patterns that do not recede to the VP. The only partially independent evidence is TMM17-Test (Table 4), where R-VPD is second best, not superior; the p=0.003 vs 'threshold 0.001' argument appears post-hoc, since under the conventional 0.05 threshold NeurVPS is significantly better.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes R-VPD, an unsupervised vanishing point detector that augments explicit line segments with implicit lines obtained from corresponding SIFT keypoints across recurring pattern instances. It contributes two benchmarks, RPVP-Synthetic (3,200 Blender-rendered images) and RPVP-Real (1,400 annotated images), and reports comparisons against LSD, J-Linkage, NeurVPS, and GPVPD using angle-accuracy AUC values and p-values. The central claims are that R-VPD outperforms all methods on synthetic images, outperforms classical methods and matches supervised deep learning methods on real-world images.","tokens_in":16305,"tokens_out":4241,"duration_ms":43191,"significance":"The core idea is original and potentially useful: detecting vanishing points from recurring patterns without training data and without relying on explicit straight lines. If the results generalize, the contribution is significant for scenes with repeated patterns and few line segments. The two RPVP datasets are a further contribution, and the authors state that code and data will be released. The stress test in Appendix D is a useful robustness check. However, the benchmark design and the statistical reporting currently leave the main generalization claim under-supported: the synthetic data are generated from exactly the geometric prior the method assumes, and the real-world ground truth is annotated by drawing lines through recurring patterns, which is the same construction R-VPD is designed to recover.","major_comments":[{"comment":"The synthetic benchmark is generated from the method's own prior. Objects are placed along L(t)=O+tD and oriented toward the vanishing point, so every recurring pattern instance forms a receding line with monotonically decreasing projected scale. This is exactly the assumption used in Section 3.4, where 'vectors from larger to smaller scale features' are regarded as pointing toward the vanishing point. Consequently, Table 1 demonstrates that R-VPD works when its prior is exactly satisfied, but it does not establish performance on images where recurring patterns do not recede to the VP, such as fronto-parallel brick or tile patterns. I request an evaluation on an independent benchmark, or on a constructed control set where repeated instances are arranged without a common vanishing line, with VP accuracy reported in both settings.","section":"Section 4.1.1 and Appendix B.1.1 versus Section 3.4"},{"comment":"The p-value test is under-specified. The manuscript does not state what test statistic is used, whether the comparison is paired across images, how the entire AUC curve is reduced to a single test, or how multiple comparisons across the eight methods are controlled. Without this information, statements such as 'p-value << 0.001' in Section 5.1 and the claim in Section 5.3 that p=0.003 (>0.001) means NeurVPS does not significantly outperform R-VPD are not verifiable. Please define the test, justify the 0.001 threshold, and report effect sizes or confidence intervals in addition to p-values.","section":"Section 5.1"},{"comment":"The angle accuracy metric uses an arbitrary nominal focal length f=(W+H)/4 for computing VP unit vectors. Since all angular errors depend on f, method rankings can change under different focal-length assumptions. Please provide a sensitivity analysis across plausible values of f, or use ground-truth camera intrinsics on RPVP-Synthetic where they are available, and state clearly whether any of the reported conclusions are robust to this choice.","section":"Section 4.2"},{"comment":"The ground truth for RPVP-Real is computed as the intersection of lines that are manually drawn through recurring patterns. This is the same implicit-line construction that R-VPD is designed to recover, which makes the real-world benchmark partially self-confirming. I ask for an assessment of annotation variability, a comparison with an independent ground-truth source (for example, vanishing points derived from 3D structure or from line annotations not tied to recurrence), and a quantification of how strongly the reported rankings depend on the annotation convention.","section":"Section 4.1.2 and Appendix B.2"}],"minor_comments":[{"comment":"The composite score SC = SL × exp(SA + SS)/N^2 mixes quantities with different units (pixels, radians, and dimensionless values), and the paper does not explain why the exponential of the summed angle and scale scores is preferable to a weighted sum or product. A brief rationale or an ablation would clarify the design.","section":"Section 3.3"},{"comment":"There are inconsistencies in the reported processing times: Appendix C states 0.86 and 0.48 seconds for NeurVPS and GPVPD on real-world images and 0.86 and 0.58 seconds on synthetic images, while Table 5 lists 0.86 and 0.52 seconds for RPVP-Real and 0.52 and 0.52 seconds for RPVP-Synth. Please correct these numbers.","section":"Table 5 and Appendix C"},{"comment":"The caption labels two stages with the same letter, '(c) Line Fit' and '(c) Inliers', which makes the pipeline stages hard to follow. Please renumber the subfigures.","section":"Figure 17 caption"},{"comment":"The closest prior work on using repeated patterns for geometric rectification, namely the methods of Pritts et al. [34-36], is cited but not included as a baseline. A comparison with at least one such method would help position R-VPD relative to existing recurrence-based geometric estimation approaches.","section":"Section 2 and Section 4.2"}],"recommendation":"major_revision","confidential_remarks":"The manuscript contains an original idea and a substantial amount of experimental work, but the benchmark construction is tightly aligned with the method's core assumption, and the statistical analysis lacks the detail needed to support the headline claims. I would encourage the editor to request an independent evaluation or a control experiment before acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Fairly useful read. The core idea—forming implicit lines from recurring feature correspondences and feeding them into a weighted RANSAC alongside explicit lines—is genuinely new in the VPD literature, and it is unsupervised. I have not seen this specific combination before. The two RPVP datasets are real assets: 3,200 synthetic renderings with ground-truth camera parameters and 1,400 labeled real images. The paper is careful to split out RPVP-Real-Exclusive when comparing with NeurVPS-tmm17, which shows awareness of data leakage.\n\nThe results are plausible on their own terms. On synthetic data R-VPD wins clearly; on RPVP-Real it beats the classical baselines and is comparable to the supervised methods on the non-overlapping subset. The failure cases are shown and the limitations section is honest.\n\nNow the soft spots, in order of importance.\n\nFirst, benchmark circularity. RPVP-Synthetic was built by placing objects along L(t)=O+tD so that repeated instances form receding lines toward the VP. RPVP-Real ground truth is the intersection of manually drawn lines through recurring patterns. That is exactly the geometric prior the method assumes: corresponding features lie on lines through the VP and their SIFT scale decreases along those lines. So Tables 1–3 mostly show that R-VPD recovers the construction used to define the ground truth. This does not make the method wrong, but it does mean the headline generalization—that R-VPD works on arbitrary images with recurring patterns—is not established. A fronto-parallel brick wall, where repeated patterns do not recede, would be a fair stress test. The TMM17-Test results provide some independent evidence, and there R-VPD is second best rather than superior.\n\nSecond, the statistics are sloppy. The p-value procedure is not specified: no test statistic, no pairing, no multiple-comparison control. The claim that NeurVPS does not significantly beat R-VPD on TMM17-Test relies on a 0.001 threshold chosen after seeing p=0.003. Under the conventional 0.05 threshold, NeurVPS wins significantly. That matters for the abstract's 'on par' claim. On RPVP-Real-Exclusive the p=0.58 comparison is fine, but the TMM17-Test framing is post hoc.\n\nThird, the angle accuracy metric uses a nominal focal length f=(W+H)/4. It is applied consistently to all methods, so it is probably fair as a relative comparison, but the absolute numbers depend on an arbitrary choice. Also, no error bars are reported and the code/data link is not yet verified. The self-citation to GRASP is reasonable, since that is the actual RP detection engine.\n\nNone of this is fatal. The central idea is sound, the datasets are useful, and the method clearly has a niche: scenes with repeated textures that lack long straight edges. The paper deserves a serious referee. I would send it to review and ask for a properly specified significance test, an independent evaluation on scenes where the recurring-pattern prior does not hold, and error bars.\n\nRecommendation: accept for peer review with a request for revision.","headline":"Genuinely new unsupervised VPD idea and two useful datasets, but the benchmark leans on the method's own geometric prior and the significance claims need tighter statistics.","tokens_in":16844,"tokens_out":3686,"would_cite":true,"duration_ms":33900,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Recurring patterns alone, without straight lines or training data, can locate vanishing points in images.","keywords":["recurrence-based vanishing point detection","implicit lines","SIFT feature correspondence","unsupervised computer vision","weighted RANSAC","perspective projection","recurring pattern datasets"],"falsifier":"Render a controlled set of images exactly like RPVP-Synthetic but with the repeated objects arranged along straight 3D lines that are parallel to the image plane, so their projected size does not decrease and the projected lines do not converge; if R-VPD still reports a finite, confident vanishing point on most of these images, the scale-and-collinearity prior is generating the VP rather than measuring one.","tokens_in":15765,"feed_emoji":"📐","tokens_out":12540,"duration_ms":114044,"temperature":0.7,"pith_summary":"Vanishing-point detection usually requires either visible straight edges or a large labeled training set. This paper argues that neither is necessary: recurring instances of the same visual pattern in a single image define implicit straight lines, and the intersection of those lines locates the vanishing point. The proposed R-VPD pipeline discovers corresponding SIFT keypoints across pattern instances, fits lines through them with a scale-based orientation pointing toward the vanishing point, and merges those implicit lines with explicit line segments through a weighted RANSAC vote. On the paper's new 3,200-image synthetic benchmark and its 1,400-image real-world benchmark, this unsupervised method beats classical line-based detectors, and on real recurring-pattern images it matches a supervised deep-learning network. The broader point a reader should take away is that perspective information is carried by texture repetition, so the bottleneck for vanishing-point detection is not only edges or labels but the ability to see what repeats.","feed_headline":"Detect vanishing points using repeated patterns, no training labels","feed_subtitle":"Matched repeated objects supply the perspective lines edge detectors miss, on par with supervised networks on real photos.","key_machinery":"The load-bearing object is the implicit line: a straight line fit to corresponding SIFT keypoints that repeat across two or more instances of a recurring pattern, with direction fixed by keypoint scale so the vector runs from larger to smaller features, i.e., toward the vanishing point. These correspondences are obtained without labeled data by hierarchical clustering of 128-dimensional SIFT descriptors, forward feature selection using linearity, angle, and scale scores (composite $S_C = S_L \\exp(S_A+S_S)/N^2$), and an optimization-inspired search over visual-word/instance groupings. The implicit lines are then fused with explicit lines in a weighted RANSAC where each line's weight starts at $w_i = \\sum_{j\\ne i} e^{-\\theta_{ij}}$ for acute angle $\\theta_{ij}$ and is updated by multiplying inliers by $\\alpha=1.2$ and outliers by $\\beta=0.8$; the winning intersection is refined by least squares and eigenvalue decomposition. This machinery is what lets a vanishing point emerge from texture repetition when no long straight edges exist.","core_discovery":"The paper's central claim is that vanishing points can be recovered from recurring patterns without any supervised training and without requiring explicit straight lines. Corresponding SIFT features across instances of a recurring pattern define implicit lines; because pattern instances project smaller as they recede, the fitted line is oriented from larger-scale to smaller-scale features, pointing toward the vanishing point. R-VPD then combines these implicit lines with any explicit line segments and uses a weighted RANSAC to vote on the intersection, refining weights so near-parallel lines dominate the consensus. The paper reports that on RPVP-Synthetic (3,200 rendered images with known camera parameters) R-VPD outperforms all four compared methods; on RPVP-Real (1,400 human-annotated images) it outperforms the classical baselines; and on the 416-image RPVP-Real-Exclusive subset it is statistically on par with NeurVPS, a supervised deep-learning method, with p = 0.58. The intended scope is single-dominant-VP natural scenes, and the paper positions the result as evidence that perspective geometry is readable from things that recur rather than only from edges.","pith_inferences":["A natural extension the paper leaves implicit is multi-VP detection: running the same implicit-line voting in a multi-model RANSAC over implicit lines could recover Manhattan or Atlanta-world vanishing points from texture repetition alone.","The scale-orientation rule could be tested independently by ablating it: randomize the SIFT scales of matched keypoints while keeping their positions; if accuracy on RPVP-Real collapses, monotone scale decrease is the real carrier of the signal, not just collinearity.","Because the real-world ground truth is drawn as lines through recurring patterns rather than through long architectural edges, an independent comparison on a dataset whose ground truth comes from explicit straight lines would show whether recurrence-derived and edge-derived vanishing points agree in general scenes.","A practical derivative is a faster, feature-light variant: replace SIFT with a cheaper keypoint detector and descriptor while keeping the weighted-RANSAC fusion, which could bring the unsupervised advantage to real-time and memory-constrained platforms."],"forward_implications":["Vanishing-point detection no longer needs scenes with long straight edges; images of periodic texture, facades, or repeated objects become usable inputs.","A no-training pipeline can reach the accuracy of a supervised deep network on real recurring-pattern imagery, so vanishing-point detection can be applied to new image domains without collecting labels.","The RPVP-Synthetic and RPVP-Real datasets give the community a benchmark where perspective information is carried by recurrence, isolating that cue from the explicit-line cue.","The recovered single vanishing point supports downstream affine rectification of one-point-perspective images, as demonstrated in the paper's frontal-view example.","Because the method's complexity is $O(n^2)$ in the number of SIFT features, scaling the descriptor extraction stage is the path to practical speed rather than dataset collection."],"supporting_citations":[{"why":"Supplies the GRASP optimization framework for discovering recurring patterns and corresponding visual words, which R-VPD adapts to produce implicit lines.","marker":"[26]"},{"why":"Provides a two-stage unsupervised RP-discovery architecture whose feature correspondences R-VPD reuses for line fitting instead of running GRASP.","marker":"[44]"},{"why":"Defines the supervised NeurVPS baseline with conic convolution and contributes the TMM17 test split used for the beyond-recurrence evaluation.","marker":"[46]"},{"why":"Defines the supervised GPVPD baseline with trainable geometric priors, the second deep-learning comparison in all benchmarks.","marker":"[25]"},{"why":"Provides the J-Linkage classical multi-structure fitting baseline that R-VPD must beat on recurring-pattern images.","marker":"[42]"},{"why":"Shows that different projected sizes of translated planar repeated elements estimate a vanishing line, the closest prior in using recurrence scale for perspective.","marker":"[34]"},{"why":"Establishes planar grouping of repeated and equally spaced structures for detecting vanishing lines and points, the theoretical ancestor of implicit-line VP detection.","marker":"[39]"},{"why":"Supplies the TMM17 natural-scene dataset and its test set, used to test R-VPD beyond images with recurring patterns.","marker":"[48]"}],"fun_headline_variants":["Vanishing points from recurring patterns, no labels needed","Unsupervised VP detection via repeated objects","Repeated patterns reveal perspective without training","Match recurring features to find vanishing points unsupervised","Recurrence-based VP detection beats classical, rivals supervised"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Recurring pattern instances lie along straight 3D lines that project through the vanishing point, and the matched SIFT keypoints shrink monotonically in scale along those lines toward the VP, so \"from larger to smaller scale\" reliably orients the implicit lines toward the VP.","fun_headline_variants_meta":{"raw":{"variants":["Vanishing points from recurring patterns, no labels needed","Unsupervised VP detection via repeated objects","Repeated patterns reveal perspective without training","Match recurring features to find vanishing points unsupervised","Recurrence-based VP detection beats classical, rivals supervised"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000178,"raw_usage":{"total_tokens":1288,"prompt_tokens":931,"completion_tokens":357,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":547,"completion_tokens_details":{"reasoning_tokens":287}},"tokens_in":547,"tokens_out":357,"duration_ms":4003,"temperature":1.0,"reasoning_tokens":287,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T23:14:00.710489+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Render a controlled set of images exactly like RPVP-Synthetic but with the repeated objects arranged along straight 3D lines that are parallel to the image plane, so their projected size does not decrease and the projected lines do not converge; if R-VPD still reports a finite, confident vanishing point on most of these images, the scale-and-collinearity prior is generating the VP rather than measuring one.","supporting_citations":[{"cited_title":"Grasp recurring patterns from a single view","cited_arxiv_id":null,"evidence_quote":"Supplies the GRASP optimization framework for discovering recurring patterns and corresponding visual words, which R-VPD adapts to produce implicit lines."},{"cited_title":"Novel 3D Scene Understanding Applications From Recurrence in a Single Image","cited_arxiv_id":"2210.07991","evidence_quote":"Provides a two-stage unsupervised RP-discovery architecture whose feature correspondences R-VPD reuses for line fitting instead of running GRASP."},{"cited_title":"Neurvps: Neural vanishing point scanning via conic convo- lution","cited_arxiv_id":null,"evidence_quote":"Defines the supervised NeurVPS baseline with conic convolution and contributes the TMM17 test split used for the beyond-recurrence evaluation."},{"cited_title":"Deep van- ishing point detection: Geometric priors make dataset vari- ations vanish","cited_arxiv_id":null,"evidence_quote":"Defines the supervised GPVPD baseline with trainable geometric priors, the second deep-learning comparison in all benchmarks."},{"cited_title":"Robust multiple struc- tures estimation with j-linkage","cited_arxiv_id":null,"evidence_quote":"Provides the J-Linkage classical multi-structure fitting baseline that R-VPD must beat on recurring-pattern images."},{"cited_title":"Rectification, and segmentation of coplanar repeated patterns","cited_arxiv_id":null,"evidence_quote":"Shows that different projected sizes of translated planar repeated elements estimate a vanishing line, the closest prior in using recurrence scale for perspective."},{"cited_title":"Planar group- ing for automatic detection of vanishing lines and points.Im- age and Vision Computing, 18(9):647–658, 2000","cited_arxiv_id":null,"evidence_quote":"Establishes planar grouping of repeated and equally spaced structures for detecting vanishing lines and points, the theoretical ancestor of implicit-line VP detection."},{"cited_title":"Detecting dominant vanishing points in natural scenes with application to composition-sensitive image retrieval","cited_arxiv_id":null,"evidence_quote":"Supplies the TMM17 natural-scene dataset and its test set, used to test R-VPD beyond images with recurring patterns."}],"review_version":1}